A functional annotation based integration of different similarity measures for gene expressions

This paper proposes a functional annotation-based method that integrates multiple gene expression similarity measures into an optimized "integrated similarity score" (ISS) to outperform individual measures in identifying similar gene pairs and predicting the functional categories of unclassified genes.

The Gist

Imagine you are a detective trying to solve a mystery in a bustling city called Genome City. In this city, there are thousands of citizens (genes). Some of these citizens have "jobs" we already know (like "makes energy" or "builds proteins"), but many others are unclassified—they have no ID cards, and we don't know what they do.

The detective's main clue is Gene Expression. Think of this as a daily diary or a music playlist for each citizen.

• If two citizens have diaries that look almost identical (they wake up at the same time, eat the same food, and go to sleep together), they are likely friends and probably do similar jobs.
• If their diaries are totally different, they probably have different lives.

The Problem: One Tool Isn't Enough

In the past, detectives tried to match these diaries using just one rule:

1. The "Shape" Detective (Correlation): Looks at the pattern of the diary. "Oh, both of them went up and down three times!"
2. The "Distance" Detective (Euclidean/Manhattan): Looks at the amount of change. "They both changed by exactly 5 units."

The problem? Sometimes the "Shape" detective is right, but the "Distance" detective is wrong, and vice versa. It's like trying to identify a song by only listening to the tempo, or only listening to the volume. You miss the whole picture.

The Solution: The "Integrated Similarity Score" (ISS)

The authors of this paper built a Super-Detective called ISS. Instead of relying on just one rule, ISS combines all the different ways of measuring similarity into one giant, super-accurate score.

But here's the tricky part: How do you decide how much to trust each rule?

• Should the "Shape" rule count for 50% and the "Distance" rule for 50%?
• Or should "Shape" count for 90%?

If you guess the wrong weights, your Super-Detective will still make mistakes.

The Secret Sauce: The "Functional Annotation" Compass

This is where the paper gets clever. The authors used a map of known jobs (called Functional Annotations) to teach the Super-Detective how to weigh the rules.

Imagine you have a group of people you know are all "Bakers."

1. You look at their diaries.
2. You calculate their similarity using the "Shape" rule and the "Distance" rule.
3. You ask: "Which rule did a better job of saying these Bakers are similar?"

If the "Shape" rule said, "These Bakers are 99% similar," but the "Distance" rule said, "They are only 10% similar," the Super-Detective learns: "Okay, for Bakers, the Shape rule is the boss. I should give the Shape rule a higher weight."

The paper created a special math formula (called FFFAG) that acts like a coach. The coach constantly tweaks the weights of the different rules, trying to minimize the difference between what the rules say and what the "Known Jobs" map says.

The Result: A Better Map

Once the Super-Detective (ISS) is trained, it creates a much better map of the city.

• Old Method: Might say Gene A and Gene B are friends.
• New Method (ISS): Says Gene A and Gene B are very close friends, and Gene C is a stranger.

The paper tested this on Yeast (a tiny fungus often used to study human biology). They found that ISS was much better at grouping genes that actually do the same job compared to the old methods.

The Grand Finale: Solving the Mystery of the Unknown

The ultimate test was to find the jobs of 40 unknown genes.

1. The team used ISS to group all the genes into clusters (like sorting people into teams based on their diaries).
2. They looked at the teams. If a team was full of "Mitochondria Workers" (genes that make energy), and one unknown gene was on that team, they guessed: "Hey, this unknown guy is probably a Mitochondria Worker too!"

The Verdict:
Using this new method, they successfully predicted the jobs of 40 unknown genes with high confidence. They even found that one unknown gene was likely involved in meiosis (how cells divide to make babies), which matched up perfectly with other scientific discoveries.

Summary Analogy

Think of the old methods as trying to identify a fruit by only looking at its color or only its weight. Sometimes you get it right, sometimes you don't.

This paper built a Smart Fruit Scanner that looks at color, weight, texture, and smell all at once. But instead of guessing how important each feature is, it looked at a basket of known fruits (apples, bananas, oranges) to learn exactly how much weight to give to "color" vs. "smell." Once trained, it could look at a mystery fruit and say, "I'm 99% sure this is a banana," even if no one had ever seen that specific banana before.

In short: They combined different ways of measuring gene activity, used known biological facts to teach the computer how to weigh those measurements, and used the result to successfully guess the jobs of genes we didn't know anything about.

Technical Summary

1. Problem Statement

Gene expression analysis relies heavily on identifying groups of genes with similar expression patterns to infer functional relationships. However, existing similarity measures (e.g., Pearson Correlation, Euclidean Distance, Spearman Rank Correlation, Manhattan Distance) have inherent limitations:

• Single-Aspect Bias: Some measures (like Euclidean Distance) reflect magnitude changes but ignore the shape of expression profiles, while others (like Pearson Correlation) are sensitive to shape but ignore magnitude.
• Lack of Biological Context: Traditional integration methods often combine these measures mathematically without incorporating biological knowledge (functional annotations) to determine the optimal weighting.
• Suboptimal Clustering: Consequently, clustering based on single or unweighted measures often fails to group genes that are functionally related, limiting the accuracy of gene function prediction for unclassified genes.

The authors aim to develop a framework that integrates multiple similarity measures into a single, superior score by leveraging biological functional annotations to determine the optimal weights.

2. Methodology

The proposed framework, called Integrated Similarity Score (ISS), operates through a three-step process:

A. Data Preprocessing and Rescoring

1. Similarity Calculation: Pairwise similarities are calculated for gene expression data using four standard measures: Euclidean Distance (ED), Pearson Correlation (PC), Spearman Rank Correlation (SRC), and Manhattan Distance (MD).
2. Unified Framework (PPV): To compare these diverse measures on a common scale, they are converted into Positive Predictive Values (PPV). This is done using Saccharomyces cerevisiae (Yeast) Gene Ontology (GO) annotations (specifically GO-Slim biological processes).
   • A gene pair is considered a "True Positive" if both genes share the same GO-Slim process annotation.
   • PPV is calculated as the ratio of gene pairs with shared annotations to the total number of gene pairs at a specific similarity threshold.

B. Weighted Integration (ISS)

The ISS is defined as a weighted linear combination of the PPVs of the individual similarity measures:
$$I_{X,Y} = \frac{\sum_{l=1}^{m} (w_l \times S_l)}{\sum_{l=1}^{m} w_l}$$
Where $S_l$ is the PPV of the $l$-th similarity measure and $w_l$ is its corresponding weight.

C. Optimization via Fitness Function (FFFAG)

To determine the optimal weights ($w_l$), the authors introduce a new fitness function called Fitness Function using Functional Annotation of Genes (FFFAG).

• Objective: Minimize the absolute difference between the Functional Similarity ($M_{ij}$) and the Integrated Similarity Score ($I_{ij}$) for all gene pairs.
• Definition: $FFFAG = \sum \sum |M_{ij} - I_{ij}|$
  • $M_{ij} = 1$ if genes $i$ and $j$ share a GO category; otherwise $0$.
  • The optimization iteratively adjusts weights (ranging from 0.1 to 1.0) to minimize this difference. If functional similarity is high (1) but expression similarity is low, the weights are adjusted to increase the ISS, and vice versa.

D. Modified TMJ

The authors also modified an existing measure, TMJ (Triangle $\times$ Jaccard), by applying the same FFFAG-based weighting strategy to create a Modified TMJ (MTMJ), allowing for a fair comparison with the proposed ISS.

3. Key Contributions

1. Novel Integration Framework (ISS): The first method to integrate different expression similarity measures using a linear combination where weights are systematically derived from biological functional annotations.
2. FFFAG Fitness Function: A new optimization metric that explicitly minimizes the discrepancy between functional similarity (ground truth from annotations) and expression similarity, ensuring the resulting score reflects biological reality.
3. Unified Scoring: Converting diverse distance/correlation metrics into a single PPV framework allows for direct comparison and integration.
4. Modified TMJ: Adapting the existing TMJ measure to incorporate functional knowledge, demonstrating that annotation-driven weighting improves even established metrics.

4. Experimental Results

The method was evaluated on six Saccharomyces cerevisiae datasets (All Yeast, Diauxic Shift, Cell Cycle, Sporulation, etc.) containing 6,072 genes.

• Performance Metrics:

  • PPV vs. Similarity Value: The ISS curve consistently outperformed individual measures (MD, ED, PC, SRC) and the original TMJ. For example, at a similarity value of 0.85 in the "All Yeast" dataset, ISS achieved a PPV of 0.92, compared to 0.68 for PC and 0.35 for MD.
  • PPV vs. Top Gene Pairs: ISS maintained the highest PPV across the top $N$ gene pairs, indicating superior ability to identify functionally linked genes at the top of the ranking.
  • Cross-Validation: 5-fold cross-validation confirmed that ISS generalizes well and consistently outperforms existing measures.

• Function Prediction:

  • Using k-medoids clustering on the ISS scores, the authors identified 12 clusters with significant functional enrichment ($p < 10^{-10}$).
  • They successfully predicted the functions of 40 unclassified yeast genes.
  • Validation:
    • YLR204W was predicted as "mitochondrion," which aligns with literature showing its role in COX1 RNA processing.
    • YDR374C was predicted as "meiosis," consistent with its known role in sporulation.
    • YOR258W was predicted as "protein folding," supported by its homology to Aprataxin.

5. Significance

• Improved Gene Function Prediction: By integrating magnitude and shape similarities weighted by biological truth, ISS provides a more accurate metric for clustering genes. This directly translates to higher accuracy in predicting the functions of uncharacterized genes.
• Biological Relevance: The approach moves beyond purely mathematical clustering by embedding biological knowledge (GO annotations) directly into the weight optimization process.
• Generalizability: While tested on Yeast, the framework of using functional annotations to optimize similarity weights is applicable to other organisms and potentially other biological data types.
• Future Directions: The authors suggest that replacing k-medoids with fuzzy clustering could further improve results, allowing genes to belong to multiple functional clusters with varying degrees of membership.

In conclusion, the paper demonstrates that a biologically informed, weighted integration of similarity measures (ISS) significantly outperforms traditional individual measures and existing integrated methods in identifying functionally related gene pairs and predicting gene functions.